{"id":"ac01fc8a-efe3-4015-b846-b8c3aecd117c","arxiv_id":"1908.05752","paper_version":6,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":8.0,"correctness_risk":"low","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper proves the inconsistency of isotonic regression at the boundary, derives consistent boundary-corrected estimators and a trimmed wild bootstrap, and applies them to monotone regression discontinuity designs.","lead":"This paper shows that isotonic regression is unreliable at the edge of its data range and builds a monotone regression discontinuity estimator that fixes this with boundary corrections and a trimmed wild bootstrap. In small samples the new estimator can cut mean squared error several-fold, and it also supplies a tuning-free lower bound for causal effects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No significant mathematical objection identified; the cube-root limit's m'(0)>0 condition is a scope restriction rather than a flaw.","rationale":"The reader and I both identify the m'(0)>0 assumption as the weakest point of the paper. I agree that the cube-root limiting distribution in Theorem 2.1(ii) and Theorem 3.1 is not uniform over flat regressions: if m'(0)=0, the quadratic drift disappears and the limit becomes the Brownian-motion GCM rather than the stated process. This is a legitimate scope condition, but not an internal inconsistency, because the paper explicitly assumes positive one-sided derivatives. Moreover, the paper's inference recommendation uses a=1/2 and the trimmed wild bootstrap, whose validity does not require m'(0)>0 and only uses a Hölder condition with γ>1/2. Thus the practical impact of the flat case is limited to the point-estimation asymptotic distribution, while consistency and the proposed confidence intervals remain defensible. I also checked the proof structure: the argmax continuous mapping arguments, the peeling tightness device, and the bootstrap multiplier CLT are coherent. The typo in Theorem 2.2 (the second probability should be unconditional, not Pr^*) is visible in the statement but the proof clearly targets the intended result. Proposition SM.2.1 assumes global continuity while Theorem 2.2 only assumes local Hölder continuity; this is a minor proof gap that can be closed because the relevant functions are only evaluated on a shrinking neighborhood of zero. These issues do not change the reader's conditional verdict, which is driven by the absence of code and data artifacts for exact reproduction.","tokens_in":38927,"tokens_out":36218,"duration_ms":350944,"concrete_test":"Run a Monte Carlo with the sharp RDD DGP m(x)=x^2 (so m'(0)=0), true jump θ=1, n=1000, a=1/3, c=1, and compare the empirical distribution of n^{1/3}(hat θ-θ) to the Theorem 3.1 density and to the no-quadratic Brownian-GCM difference. If the latter matches, the m'(0)>0 condition is necessary for the stated cube-root law. Then repeat with a=1/2 and the trimmed wild bootstrap; if coverage is near nominal, the inference recommendation is robust to a flat conditional mean near the cutoff.","verdict_should_be":"UNCHANGED","load_bearing_attack":"No significant mathematical objection identified. The main theorems are internally consistent under Assumptions 2.1 and 3.1, and I find no circular step or derivation gap that invalidates the central claims. The most load-bearing scope condition is Assumption 2.1(iv)/3.1(iv): the cube-root limits in Theorem 2.1(ii) at a=1/3 and Theorem 3.1 require m'(0)>0 so that the quadratic drift appears in the limiting greatest convex minorant. If m'(0)=0, the drift term vanishes and the limit changes to the Brownian-motion GCM of the a>1/3 case; the paper does not cover this case. This is a real limitation of the stated point-estimation distribution, but it does not threaten consistency, and the bootstrap inference is deliberately run at a=1/2, where the m'-term is absent and only a Hölder condition is required. Separately, Theorem 2.2 as written compares a bootstrap CDF to Pr^*(...) of a nonrandom event; the second probability should be unconditional. The proof establishes the intended bootstrap consistency, so this is a statement-level typo, not a mathematical failure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops asymptotic theory for the isotonic regression estimator evaluated at boundary points of the covariate support, motivated by monotone regression discontinuity designs (iRDD). It shows that the natural estimator evaluated at the closest boundary observation is inconsistent, introduces boundary-corrected estimators evaluated at c n^{-a}, and derives their limit distributions for slow, cube-root, and fast correction rates (Theorem 2.1). It further proposes a trimmed wild bootstrap for the fast-correction regime and proves its consistency (Theorem 2.2), with sharp and fuzzy RDD analogues (Theorems 3.1-3.3). Monte Carlo experiments show large MSE improvements over local polynomial and k-NN estimators, and the method is applied to estimate the incumbency advantage in U.S. House elections. The proofs are detailed and, apart from the issues listed below, internally consistent.","tokens_in":39145,"tokens_out":16115,"duration_ms":142445,"significance":"If the stated results hold after the corrections requested below, this is a substantial contribution to both shape-constrained regression and regression discontinuity methodology. The paper provides a self-contained derivation of boundary limit theory for isotonic regression, including a novel treatment of tightness without strong approximation, and demonstrates that a wild bootstrap works at the boundary without subsampling or smoothing—a property known to fail at interior points. The explicit limit distributions, the consistent bootstrap procedure, and the detailed Monte Carlo evidence make the results immediately usable in practice. The empirical application to incumbency effects illustrates the method's relevance. The proofs are rigorous and the derivation is not circular; the main weakness is that one bootstrap theorem is misstated and another central theorem is stated without proof.","major_comments":[{"comment":"The displayed statement compares two bootstrap probabilities, but the second term, Pr^*( n^{(1-a)/2}(\\hat m(cn^{-a}) - m(0)) \\le u ), is a random indicator conditional on the data because the event inside is measurable with respect to the original sample. As printed, the theorem is therefore false; it should be the unconditional probability Pr( n^{(1-a)/2}(\\hat m(cn^{-a}) - m(0)) \\le u ). The proof in the appendix actually establishes convergence of the bootstrap CDF to Pr( DL_{[0,\\infty)}(\\sqrt{\\sigma^2(0)/(c f(0))} W_t )(1) \\le u ) and then invokes Theorem 2.1(ii) for the unconditional limit, so the intended statement is clear and the correction is a single symbol. Note that Theorem 3.3 states the analogous result correctly with an unconditional probability, confirming the typo.","section":"Section 2.3, Theorem 2.2"},{"comment":"Theorem 3.3 is a central result for inference in the sharp iRDD and is used in the empirical application, but no proof is provided in the Appendix or Supplementary Material. The Appendix's \"Proofs of main results\" covers Theorems 2.1, 2.2, 3.1, 3.2, and Remark 2.1 only; Theorem 3.3 is stated and then the section moves on. Since the theorem concerns the bootstrap distribution of a difference of two boundary-corrected estimators, a proof or a detailed derivation from Theorem 2.2 together with the independence of the two sides must be added.","section":"Section 3.4, Theorem 3.3"},{"comment":"The fuzzy RDD theorem does not explicitly require p_+ \\neq p_- (or p_+ > p_-), which is necessary for the denominator p_+ - p_- in the limit expression and for the fuzzy estimator to be well defined. The assumptions list p \\in M[-1,1], p'_\\pm > 0, and p_\\pm \\in (0,1), but a nondecreasing p can satisfy these with p_+ = p_-. The identification condition from Section 3.1, namely the discontinuity in the treatment assignment probability, should be added as an explicit assumption in Theorem 3.2.","section":"Section 3.3, Theorem 3.2"}],"minor_comments":[{"comment":"The cube-root boundary limit at a = 1/3 requires m'(0) > 0, so the flat case m'(0) = 0 is outside the scope of the stated distribution. Since the paper recommends a = 1/3 for point estimation, a remark explaining that the limit changes when the derivative vanishes and that the method is not designed for that case would help readers assess applicability.","section":"Section 2.2, Assumption 2.1(iv)"},{"comment":"At the recommended rule-of-thumb c = 1, the empirical coverage is approximately 0.85-0.91 for a nominal 95% level across n = 200, 500, 1000, which is a noticeable finite-sample undercoverage. The paper reports these numbers and notes that coverage improves with larger c, but given that c = 1 is the default recommendation used in the empirical section, a more explicit acknowledgment of this undercoverage would be appropriate.","section":"Table SM.4"},{"comment":"The notation \\check\\theta and \\check\\theta^* is introduced in the text preceding the theorem, but the theorem statement should either define these quantities or reference the paragraph where they are defined, to avoid ambiguity for a reader who goes directly to the theorem.","section":"Section 3.4, Theorem 3.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid contribution with a sound core derivation, detailed proofs, and useful Monte Carlo evidence. The required revision is mostly about completing and correcting the presentation of two key theorems: the typo in Theorem 2.2 (the second probability should be unconditional) and the missing proof of Theorem 3.3. The omitted positivity condition in Theorem 3.2 should also be added. These are all fixable within the manuscript's scope, so I view the paper as a promising candidate after revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know: this is the missing boundary theory for isotonic regression in RDDs, and I think the main claims are correct. The paper proves the raw isotonic estimator is inconsistent at the boundary under random design, then constructs boundary-corrected estimators with a cube-root limit at a=1/3 when m'(0)>0 and a slower n^{-(1-a)/2} limit under only Holder smoothness. The trimmed wild bootstrap is genuinely new: it restores bootstrap consistency at the boundary without subsampling or smoothing.\n\nWhat's new: the boundary behavior with random design is not in Anevski-Hossjer, and the one-sided GCM slope argument explaining the always-below-boundary property is neat. The proofs are detailed and internally consistent; I found no circular step. The Monte Carlo results show large MSE reductions over local polynomial and k-NN under monotone designs, and the incumbency application is sensible.\n\nSoft spots: the cube-root distribution (Theorems 2.1(ii) and 3.1) requires m'(0)>0. If the regression is flat at the cutoff, the quadratic drift vanishes and the stated limit no longer applies; the paper doesn't cover that case. That's a scope restriction, not a fatal flaw, but it's load-bearing for point estimation. Also, the bootstrap coverage in Table SM.4 is about 90% for nominal 95% at the recommended c=1; larger c helps but isn't recommended. The MSE-optimal c underperforms rule-of-thumb c=1 in small samples, so the tuning advice is pragmatic but not clearly optimal. My main practical condition: the authors don't provide code or data, so replicating the Monte Carlo requires weeks of manual implementation. Minor: Theorem 2.2's second probability should be unconditional, not Pr^*, but the proof makes the intended bootstrap consistency clear.\n\nBottom line: the central argument is sound. I'd send this to a serious referee, with a request for code and data and a note about the m'(0)>0 caveat. Worth engaging with.","headline":"Solid boundary asymptotic theory for isotonic regression in RDDs; the main claims hold up, with a scope restriction on m'(0)>0 and a minor typo in Theorem 2.2, and it deserves peer review.","tokens_in":39660,"tokens_out":2312,"would_cite":true,"duration_ms":21504,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["62G05","62G08","62G20","62P20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Boundary-corrected isotonic regression is consistent at regression-discontinuity cutoffs, and a trimmed wild bootstrap makes inference work without subsampling or smoothing.","keywords":["regression discontinuity designs","shape constraints","monotonicity","isotonic regression","boundary point","wild bootstrap","greatest convex minorant","cube-root asymptotics"],"falsifier":"Generate data with $Y = 1 + \\varepsilon$, $X$ uniform on $[0,1]$, and $\\varepsilon$ i.i.d. standard normal, so $m'(0)=0$. Compute $n^{1/3}(\\hat m(c n^{-1/3}) - 1)$ for large $n$ across many replications and compare its distribution with the theorem's Brownian-plus-parabola GCM limit: with no quadratic drift the observed distribution must diverge from that limit, confirming the strict-slope assumption is load-bearing.","tokens_in":38726,"feed_emoji":"📈","tokens_out":12431,"duration_ms":107079,"temperature":0.7,"pith_summary":"This paper shows that the natural isotonic regression at the boundary of the covariate support—the setting regression discontinuity designs require—is inconsistent: evaluating the fit at the closest observation underestimates the boundary limit with probability tending to one. Evaluating the isotonic fit at a point that shrinks toward the boundary as $c n^{-a}$ repairs this, yielding consistency at rate $n^{(1-a)/2}$ for $a \\in [1/3,1)$, and at $a=1/3$ it recovers the classical cube-root non-normal limit built from greatest convex minorants of Brownian motion. For faster corrections, the paper's trimmed wild bootstrap is consistent without subsampling or additional smoothing, a property that fails at interior points. Applied to sharp and fuzzy regression discontinuity designs, the paper derives limiting distributions for the causal effect and reports large finite-sample MSE reductions relative to unrestricted local polynomial and $k$-nearest-neighbor estimators under monotone designs. The empirical illustration re-estimates the U.S. House incumbency advantage at 13.8 percentage points with a 95% confidence interval of [6.6%, 26.5%].","feed_headline":"Monotone RDD estimator fixes boundary bias at cube-root speed","feed_subtitle":"A trimmed wild bootstrap gives valid confidence intervals without smoothing or subsampling.","key_machinery":"The load-bearing object is the greatest convex minorant (GCM) of the cumulative sum diagram $(F_n, M_n)$, whose left derivative at any point equals the isotonic regression estimate. Boundary correction evaluates that derivative at $x = c n^{-a}$, a point that approaches zero fast enough to remove the extreme-value bias but slowly enough that local observations accumulate. The quadratic drift $t^2 c m'(0)/2$ in the limiting Brownian-plus-parabola process comes from the regression slope at the boundary and disappears for $a > 1/3$, which is exactly why the trimmed wild bootstrap can be consistent: trimming removes the estimated quadratic term that the bootstrap cannot reproduce.","core_discovery":"The central discovery is that the boundary-corrected isotonic estimator $\\hat m(c n^{-a})$ is consistent for $m(0)$ and, for $a \\in [1/3,1)$, converges at rate $n^{(1-a)/2}$ to a functional of the greatest convex minorant of Brownian motion. At $a=1/3$ the rate is $n^{1/3}$ and the limit is the left derivative at 1 of the greatest convex minorant of $\\sqrt{\\sigma^2(0)/(c f(0))}\\,W_t + (t^2 c/2) m'(0)$, the boundary analogue of the interior-point cube-root law. For $a \\in (1/3,1)$ the same limit holds under only $\\gamma$-Hölder smoothness, and the trimmed wild bootstrap—which replaces the fitted curve on $[0, c n^{-a}]$ by its boundary value before generating multiplier residuals—is consistent without subsampling or smoothing. In the sharp RDD, $n^{1/3}(\\hat\\theta - \\theta)$ converges to the difference of two independent such slope functionals on either side of the cutoff; in fuzzy designs a joint limit with isotonic treatment-probability fits applies. The paper also proves that the uncorrected estimator at the extreme observation $X_{(1)}$ is inconsistent and systematically understates $m(0)$.","pith_inferences":["The same boundary-correction-plus-trimming recipe should transfer to other shape-constrained estimators whose limits are GCM functionals, such as monotone density estimation at zero, where bootstrap consistency is otherwise delicate.","Because the uncorrected boundary estimate is biased toward zero effect, one-sided inference on it could be made entirely tuning-free; the paper notes the lower-bound property but does not develop that inference.","The simulation pattern in which larger $c$ raises bootstrap coverage suggests a data-driven coverage-optimal $c$ could close the remaining gap between nominal and empirical coverage; the paper leaves that choice open.","A flat regression at the cutoff, where $m'(0)=0$, is excluded by the point-estimation theory; testing whether a pre-test for positive slope can select between the $a=1/3$ and $a>1/3$ regimes would be a direct follow-up."],"forward_implications":["At $a = 1/3$, the boundary-corrected isotonic estimator achieves the same $n^{1/3}$ rate and non-normal GCM limit as the interior-point estimator, but for the boundary parameter $m(0)$.","For $a \\in (1/3,1)$, valid confidence intervals follow from the trimmed wild bootstrap without subsampling or nonparametric smoothing; the rule-of-thumb $a = 1/2$ gives $n^{-1/4}$ rate.","The sharp RDD limit is the difference of two independent GCM slope functionals, giving a formal asymptotic distribution for monotone sharp designs.","The uncorrected boundary estimator is inconsistent and tends to understate $m(0)$, so a tuning-free isotonic RDD estimate acts as a lower bound on the causal effect.","Simulations across monotone designs, including heteroskedastic and steep cases, show large MSE reductions for the iRDD estimator relative to local polynomial and $k$-nearest-neighbor estimators."],"supporting_citations":[{"why":"Establishes the interior-point cube-root limiting distribution for isotonic regression that the boundary case with $a<1/3$ matches.","marker":"Wright (1981)"},{"why":"Provides the earlier inconsistency result for isotonic regression at a discontinuity in fixed-design settings; the paper extends this to random covariates.","marker":"Anevski and H¨ossjer (2002)"},{"why":"Supplies the argmax continuous mapping theorem and tightness framework used to prove the boundary limit distributions.","marker":"Kim and Pollard (1990)"},{"why":"Supplies empirical-process Donsker and argmax theorems used to establish weak convergence of localized processes.","marker":"van der Vaart and Wellner (1996)"},{"why":"Analyzes boundary behavior of the Grenander estimator and uses strong approximation, an approach the paper avoids with a partitioning argument.","marker":"Kulikov and Lopuha¨a (2006)"},{"why":"Gives absolute continuity of the GCM-of-Brownian-motion limiting distribution, used to justify the bootstrap approximation.","marker":"Groeneboom (1983)"},{"why":"Defines the sharp and fuzzy RDD identification and estimation framework into which the iRDD estimators plug.","marker":"Hahn et al. (2001)"},{"why":"Supplies the U.S. House elections data and incumbency benchmark that the empirical illustration re-estimates.","marker":"Lee (2008)"},{"why":"Documents bootstrap failure for the Grenander estimator at interior points, motivating why boundary trimming can restore consistency.","marker":"Sen et al. (2010)"},{"why":"Offers generic bootstrap remedies for cube-root estimators; the paper shows boundary correction alone suffices here.","marker":"Cattaneo et al. (2020b)"}],"fun_headline_variants":["Boundary-corrected isotonic regression hits cube-root rate","Isotonic RDD fix: consistency at the boundary","Bootstrap without smoothing for boundary isotonic estimates","Naive isotonic fails at boundary; corrected version works","Cube-root limit for isotonic regression at the cutoff"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The recommended cube-root point estimator and its stated limiting distribution require the conditional mean to be strictly increasing at the cutoff ($m'(0)>0$); if the mean is flat near the cutoff, the quadratic drift in the limiting process disappears and the stated result no longer applies.","fun_headline_variants_meta":{"raw":{"variants":["Boundary-corrected isotonic regression hits cube-root rate","Isotonic RDD fix: consistency at the boundary","Bootstrap without smoothing for boundary isotonic estimates","Naive isotonic fails at boundary; corrected version works","Cube-root limit for isotonic regression at the cutoff"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000407,"raw_usage":{"total_tokens":2104,"prompt_tokens":926,"completion_tokens":1178,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":1099}},"tokens_in":542,"tokens_out":1178,"duration_ms":10018,"temperature":1.0,"reasoning_tokens":1099,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:08:18.317155+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate data with $Y = 1 + \\varepsilon$, $X$ uniform on $[0,1]$, and $\\varepsilon$ i.i.d. standard normal, so $m'(0)=0$. Compute $n^{1/3}(\\hat m(c n^{-1/3}) - 1)$ for large $n$ across many replications and compare its distribution with the theorem's Brownian-plus-parabola GCM limit: with no quadratic drift the observed distribution must diverge from that limit, confirming the strict-slope assumption is load-bearing.","supporting_citations":[],"review_version":1}